Attura AI · Anomaly Detection
Attura AI uses Isolation Forest — a label-free anomaly detection method — to flag attendance, overtime, leave, and payroll patterns that deviate from your organization's norms.
The AI points. HR decides.
Coming Soon · In developmentYour HRIS records attendance, overtime, leave, business trips, and payroll — thousands of new rows every month.
The problem is not a lack of data. Unusual patterns can look small when seen one record at a time, and manual review only gets harder as volume grows.
Anomaly detection looks across records at once, then hands the result to a human.
An anomaly is simply an observation that does not follow the pattern of the others.
A 13:47 clock-in among a row of 08:00s does not mean a violation. It is a signal to review — maybe the shift changed, maybe a half-day permission, maybe it does need checking. The decision stays with HR.
How it works
Isolation Forest does not look for "what is wrong". It draws separating lines at random positions, then counts: how many lines does it take before an observation stands alone?
Each point is one observation from HRIS data — for example, an employee's overtime pattern in one period.
A feature is picked at random, then a split value is picked at random between its smallest and largest values. The space splits in two.
The splitting repeats. Points that crowd together are hard to separate — they keep sharing space with their neighbours.
An unusual observation gets boxed in alone after only a few lines. That number of lines is its "path length".
The process repeats across many random trees. A short average path yields a high anomaly signal — ready for HR review.
9+ lines to separate it
It keeps sharing space with many neighbours. Its path is long — its anomaly signal is low.
3 lines and it stands alone
Isolated quickly in nearly every random tree. Its path is short — its anomaly signal is high.
This intuition — yang tidak biasa lebih mudah diisolasi — adalah seluruh dasar metode ini. Tanpa label, tanpa daftar aturan, tanpa definisi "salah".
The same model reads four data domains already in Attura. Always in the same shape: pattern → signal → HR review.
An anomaly does not mean fraud, does not mean an employee is at fault, and does not mean payroll must be wrong. An anomaly is a signal to review. The numbers in the visuals above are illustrative examples.
The AI does not make HR decisions. It does not declare anyone at fault. It does not replace human judgment. It helps prioritise what deserves a look first.
Model
Anomaly signal · overtime pattern · July period
The model does not read human stories. It reads patterns in the data it is given.
HR
The signal opens the context — the context produces the decision.
Isolation Forest (Liu, Ting & Zhou, IEEE ICDM 2008) is an unsupervised anomaly detection method. Each isolation tree is built by picking a feature at random, then a random split value between that feature's minimum and maximum — repeatedly, until observations are isolated.
The number of splits needed to isolate an observation equals the path length from the tree's root to its leaf. Averaged over many random trees (an ensemble), a shorter path means the observation is easier to isolate — and more likely to be an anomaly.
The resulting score is a ranking of anomaly signals, not a probability of wrongdoing. Thresholds and interpretation need to be contextualised to each organization's data.
Reference: F. T. Liu, K. M. Ting, Z.-H. Zhou, "Isolation Forest", IEEE ICDM 2008 · scikit-learn documentation: outlier detection.
Attura AI is under development and will be available automatically to Advance plan subscribers once released — reading the data already in your HRIS.